Embed Size (px)
Citation preview
Ordered Ramsey numbers
David Conlon∗ Jacob Fox† Choongbum Lee‡ Benny Sudakov§
Abstract
Given a labeled graph H with vertex set 1, 2, . . . , n, the ordered Ramsey number r<(H) is
the minimum N such that every two-coloring of the edges of the complete graph on 1, 2, . . . , Ncontains a copy of H with vertices appearing in the same order as in H. The ordered Ramsey
number of a labeled graph H is at least the Ramsey number r(H) and the two coincide for
complete graphs. However, we prove that even for matchings there are labelings where the
ordered Ramsey number is superpolynomial in the number of vertices. Among other results, we
also prove a general upper bound on ordered Ramsey numbers which implies that there exists
a constant c such that r<(H) ≤ r(H)c log2 n for any labeled graph H on vertex set 1, 2, . . . , n.
1 Introduction
Given a graph H, the Ramsey number r(H) is defined to be the smallest natural number N such that
every two-coloring of the edges of KN contains a monochromatic copy of H. That these numbers
exist was first proved by Ramsey [34] and rediscovered independently by Erdos and Szekeres [14].
The most famous question in graph Ramsey theory is that of estimating the Ramsey number
r(Kn) of the complete graph Kn on n vertices. However, despite some smaller order improve-
ments [9, 35], the standard estimates [13, 14] that 2n/2 ≤ r(Kn) ≤ 22n have remained largely
unchanged for nearly seventy years. After the complete graph, the next most classical topic in the
area is the study of Ramsey numbers of sparse graphs, that is, graphs with certain upper bound
constraints on the degrees of their vertices. This direction was pioneered by Burr and Erdos [6] in
1975 and the topic has since played a central role in graph Ramsey theory.
Answering a question of Burr and Erdos, Chvatal, Rodl, Szemeredi and Trotter [7] proved that
for every ∆ there is c(∆) such that every graph H on at most n vertices with maximum degree
∆ satisfies r(H) ≤ c(∆)n. That is, Ramsey numbers of bounded-degree graphs grow linearly in
the number of vertices. Another stronger conjecture of Burr and Erdos remained open until very
recently. We say that a graph is d-degenerate if every subgraph has a vertex of degree at most d.
∗Mathematical Institute, Oxford OX2 6GG, United Kingdom. Email: [email protected]. Research
supported by a Royal Society University Research Fellowship.†Department of Mathematics, Stanford University, Stanford, CA 94305. Email: [email protected]. Re-
search supported by a Packard Fellowship, by NSF Career Award DMS-1352121 and by an Alfred P. Sloan Fellowship.‡Department of Mathematics, MIT, Cambridge, MA 02139-4307. Email: cb [email protected]. Research sup-
ported by NSF Grant DMS-1362326§Department of Mathematics, ETH, 8092 Zurich, Switzerland. Email: [email protected]. Re-
search supported in part by SNSF grant 200021-149111.
1
Equivalently, a graph is d-degenerate if there is an ordering v1, v2, . . . , vn of its vertices such that
each vertex vi has at most d neighbors vj with j < i. Burr and Erdos conjectured that for every
d there is c(d) such that every d-degenerate graph H on at most n vertices satisfies r(H) ≤ c(d)n.
Building on earlier work by several authors [2, 20, 21, 26, 27], this conjecture has now been solved
by Lee [28].
In this paper, we will study analogues of these results for ordered graphs. An ordered graph or
labeled graph H on n vertices is a graph whose vertices have been labeled with 1, 2, . . . , n. An
ordered graph G on [N ] := 1, 2, . . . , N contains an ordered graph H on [n] if there is a mapping
φ : [n] → [N ] such that φ(i) < φ(j) for 1 ≤ i < j ≤ n and (φ(i), φ(j)) is an edge of G whenever
(i, j) is an edge of H. Given an ordered graph H, the ordered Ramsey number r<(H) is defined to
be the smallest natural number N such that every two-coloring of the edges of the complete graph
on [N ] contains a monochromatic ordered copy of H. The natural analogue of this function for q
colors will be denoted by r<(H; q).
In some sense, the study of ordered Ramsey numbers is as old as Ramsey theory itself. One
of the most famous results in the classic 1935 paper of Erdos and Szekeres [14] states that every
sequence x1, x2, . . . , xr of r ≥ (n−1)2 +1 distinct real numbers contains an increasing or decreasing
subsequence of length n. To prove this, consider a red/blue-coloring of the edges of the complete
graph on vertex set [r] where (i, j) with i < j is red if xi < xj and blue otherwise. Note that a
subsequence is increasing or decreasing if and only if it forms a monotone monochromatic path in
this edge coloring. We may therefore assume that there is no monotone red path of length n. We
now label each vertex by the length of the longest monotone red path ending at that vertex. It is
easy to see that any set of vertices with the same label forms a blue clique. Therefore, since there
are only n−1 possible labels and r ≥ (n−1)2 +1 vertices, we may conclude that there exists a blue
clique of order at least n and the result of Erdos and Szekeres follows. In proving this result, we
have also shown that r<(Pn) ≤ (n− 1)2 + 1, where Pn is the monotone path with n vertices. Since
this is easily seen to be tight, we have r<(Pn) = (n− 1)2 + 1. More generally, for q colors instead
of 2, the q-color ordered Ramsey number of the monotone path satisfies r<(Pn; q) = (n− 1)q + 1.
Another foundational result in Ramsey theory, known as the happy ending theorem, also has a
natural proof using ordered Ramsey numbers. This result, again due to Erdos and Szekeres [14],
states that for each positive integer n there is an integer N such that every set of N points in the
plane in general position (that is, with no three on a line) contains n points which are the vertices
of a convex polygon. We write g(n) for the smallest such N . To estimate g(n), it will be useful
to generalize the concept of ordered Ramsey numbers to hypergraphs. That is, given an ordered
k-uniform hypergraph H and a positive integer q, we let r<(H; q) be the smallest natural number
N such that every q-coloring of the edges of the complete k-uniform hypergraph on [N ] contains a
monochromatic ordered copy of H.
Suppose now that we have N points in the plane in general position. By rotating the plane
if necessary, we may assume that no two points are on a vertical line. Denote the N points by
pi = (xi, yi) with x1 < . . . < xN . A set P of points in the plane forms a cup (respectively, cap) if the
points in P lie on the graph of a convex (respectively, concave) function. Note that both cups and
caps form convex polygons. Consider a red/blue-coloring of the edges of the complete 3-uniform
hypergraph on 1, 2, . . . , N where i, j, k with i < j < k is red if pi, pj , pk form a cup and blue
2
if pi, pj , pk form a cap. If we write P(k)n for the monotone k-uniform tight path on 1, 2, . . . , n,
where i, i+ 1, . . . , i+ k− 1 is an edge for 1 ≤ i ≤ n− k+ 1, then we see that a sequence of points
forms a cap or a cup if and only if the corresponding vertices form a monochromatic ordered copy
of P(3)n . Hence, g(n) ≤ r<(P
(3)n ), which is known to equal
(2n−4n−2
)+ 1.
An extension of the happy ending theorem proved by Pach and Toth [33] shows that for every
positive integer n there is an integer N such that every set of N convex sets in the plane in general
position contains n convex sets which are in convex position. We write h(n) for the smallest such
N . It was shown by Fox, Pach, Sudakov and Suk [18] that h(n) ≤ r<(P(3)n ; 3), leading the authors
to study the growth of ordered hypergraph Ramsey numbers for monotone paths. The results
of [18], and subsequent improvements made by Moshkovitz and Shapira [31], show that for k ≥ 3
the ordered Ramsey number r<(P(k)n ; q) grows as a (k − 2)-fold exponential in nq−1. This is in
stark contrast to the classical unordered problem, where the Ramsey number of P(k)n grows linearly
in the number of vertices for all uniformities k. More recently, ordered Ramsey numbers for tight
paths were also used by Milans, Stolee and West [30] to give bounds on the minimum number of
interval graphs whose union is the line graph of Kn.
While there has been much progress on understanding the ordered Ramsey numbers of monotone
paths, there has been surprisingly little work on more general ordered graphs and hypergraphs. In
this paper, we attempt to bridge this gap by conducting a more systematic study of ordered Ramsey
numbers, focusing on the case of graphs. We note that a similar study was conducted independently
by Balko, Cibulka, Kral and Kyncl [4] and that there are overlaps between many of our results.
One of the more striking aspects of the discussion above is the vast difference between the usual
Ramsey number and the ordered Ramsey number for monotone paths of high uniformity. Our
first result shows that a large gap exists already for graphs, even when the graph is as simple as
a matching, where the ordinary Ramsey number is clearly linear. Here and throughout the paper,
all logs are taken to base 2.
Theorem 1.1. There exists a positive constant c such that, for all even n, there exists an ordered
matching M on n vertices with
r<(M) ≥ nc logn/ log logn.
This lower bound actually holds for almost every ordering of a matching on n vertices, so that
Theorem 1.1 represents typical rather than atypical behavior. An almost matching upper bound
is provided by the following simple theorem (which, in a slightly weaker form, also follows from a
result of Cibulka, Gao, Krcal, Valla and Valtr [8]).
Theorem 1.2. For any ordered matching M on n vertices,
r<(M) ≤ ndlogne.
More generally, we can prove that there exists a constant c such that for any ordered graph H
on n vertices with degeneracy d, r<(H) ≤ ncd logn, where the degeneracy of an ordered graph H
is the smallest d for which the corresponding unordered graph is d-degenerate. This is a special
case of an even more general result. To state this result, we define the interval chromatic number
χ<(H) of an ordered graph H to be the minimum number of intervals into which the vertex set of
3
H may be partitioned so that no two vertices in the same interval are adjacent. This is similar to
the notion of chromatic number but now the independent sets must also be intervals in the given
ordering. For any graph H, it is easy to see that there is an ordering of the vertices of H such that
the interval chromatic number is the same as the chromatic number.
Interval chromatic number plays a key role in the study of extremal problems on ordered graphs
(see, for example, [32]). In particular, Pach and Tardos [32] observed that the maximum number
ex<(n,H) of edges an ordered graph on n vertices can have without containing the ordered graph
H is (1− 1
χ<(H)− 1+ o(1)
)(n
2
).
The following result shows that it also plays a fundamental role in ordered Ramsey theory.
Theorem 1.3. There exists a constant c such that for any ordered graph H on n vertices with
degeneracy d and interval chromatic number χ,
r<(H) ≤ ncd logχ.
In particular, this result implies that there are orderings under which the ordered Ramsey
number of a d-degenerate graph with n vertices is polynomial in n. The following result shows that
restricting the interval chromatic number is still not enough to force the ordered Ramsey number
to be linear, even for matchings. It is also close to best possible, since an elementary argument
(see Section 2) shows that r<(M) ≤ n2 for any matching M with n vertices and interval chromatic
number 2.
Theorem 1.4. There exists a positive constant c such that, for all even n, there exists an ordered
matching M on n vertices with interval chromatic number 2 and
r<(M) ≥ cn2
log2 n log logn.
So far, our results have focused on very sparse graphs. For denser graphs, the ordered Ramsey
number behaves more like the usual Ramsey number. Indeed, we have the following result, which
generalizes a result of the first author [10] in two ways: firstly, by estimating the ordered Ramsey
number rather than just the usual Ramsey number and, secondly, by replacing maximum degree
with degeneracy.
Theorem 1.5. There exists a constant c such that for any ordered graph H on n vertices with
degeneracy at most d,
r<(H) ≤ 2cd log2(2n/d).
This result is close to sharp when d is very large and also, by Theorem 1.1, when d is very small.
We note two simple corollaries of this theorem. The first, proved for the usual Ramsey number
in [10], says that for any graph on n vertices with o(n2) edges, the ordered Ramsey number is 2o(n).
This follows by noting that any graph on n vertices with at most δn2 edges has degeneracy at most
(2δ)1/2n and substituting into Theorem 1.5. More precisely, we have the following result.
4
Corollary 1.6. There exists a constant c such that for any ordered graph H on n vertices with at
most δn2 edges,
r<(H) ≤ 2c√δ log2(1/δ)n.
Since any graph H with degeneracy at least d contains a subgraph of minimum degree at least
d, a simple application of the probabilistic method implies that the Ramsey number of H must be
at least 2d2 . Using also the elementary observation that r(H) ≥ n, we have the following corollary.
Corollary 1.7. There exists a constant c such that for any ordered graph H on n vertices with
degeneracy d,
r<(H) ≤ r(H)cγ(H),
where γ(H) = minlog2(n/d), d log(n/d).
It is also interesting to define an off-diagonal variant of the ordered Ramsey number. Given
two ordered graphs G and H, we define the ordered Ramsey number r<(G,H) to be the smallest
natural number N such that any two-coloring of the edges of the complete graph on vertex set
[N ] in red and blue, say, contains either a red ordered copy of G or a blue ordered copy of H. To
give an example, we note that the proof of the Erdos–Szekeres theorem on monotone subsequences
given earlier actually shows that
r<(Pm,Kn) = (m− 1)(n− 1) + 1. (1)
In the classical case, the most intensively studied off-diagonal Ramsey number is r(Kn,K3).
It is easy to see that r(Kn,K3) ≤ n2 and a well-known result of Ajtai, Komlos and Szemeredi [3]
improves this to r(Kn,K3) = O( n2
logn). In a remarkable breakthrough, Kim [24] showed that this
upper bound is essentially tight, that is, that r(Kn,K3) = Ω( n2
logn). Recent advances in the study of
triangle-free processes [5, 16] have led to further improvements in these bounds, so that r(Kn,K3)
is now known up to an asymptotic factor of 4.
The main off-diagonal problem that we treat in the ordered case is that of determining the
ordered Ramsey number r<(M,K3), where M is an ordered matching. For the upper bound, we
could only prove the trivial estimate r<(M,K3) ≤ r<(Kn,K3) = O( n2
logn). For the lower bound, we
have the following result, which improves considerably on the trivial linear bound.
Theorem 1.8. There exists a positive constant c such that, for all even n, there exists an ordered
matching M on n vertices with
r<(M,K3) ≥ c( n
log n
)4/3.
In the next section, we will study the ordered Ramsey number of matchings, proving Theo-
rems 1.1, 1.2 and 1.4. We will also show that the ordered Ramsey number of matchings with
bounded bandwidth is at most polynomial in the number of vertices. In Section 3, we will prove
Theorems 1.3 and 1.5. The proof of Theorem 1.8 is given in Section 4. We conclude, in Sections 5
and 6, with some further remarks and a collection of open problems. In particular, we observe a
connection between ordered Ramsey numbers and hypergraph Ramsey numbers. We also classify
those graphs which have linear ordered Ramsey number in every ordering. Throughout the paper,
we omit floor and ceiling signs whenever they are not essential. We also do not make any serious
attempt to optimize absolute constants in our statements and proofs.
5
2 Matchings
We begin with a simple upper bound for the ordered Ramsey number of matchings in terms of the
number of vertices n and the interval chromatic number χ. By saying that the χ-partite graph
Kn′,n′,...,n′ is trivially ordered, we mean that the vertices of each partite set appear as an interval.
Theorem 2.1. Let M be an ordered matching on [n] and let K = Kn′,n′,...,n′ be a trivially ordered
χ-partite graph with χ ≥ 2. Then
r<(M,K) ≤ ndlogχen′.
In particular, if M is an ordered matching on n vertices of interval chromatic number χ, then
r<(M) ≤ ndlogχe+1.
Proof. It suffices to prove the statement when χ = 2j for positive integers j. We prove this by
induction on j. For the base case j = 1, we have to show that r<(M,Kn′,n′) ≤ nn′. To see this,
suppose that we are given a two-coloring of the edges of KN1 with N1 = nn′. Partition [N1] into n
consecutive intervals V1, V2, . . . , Vn, each of length n′. We try to embed a red copy of M by placing
vertex i in the set Vi for each i = 1, 2, . . . , n. If this procedure does not produce a red copy of M ,
then there are two indices i1 and i2 such that every edge between Vi1 and Vi2 is blue. That is, we
find either a red copy of M or a blue copy of Kn′,n′ .
Let Nj = njn′ and suppose that any two-coloring of the edges of [Nj−1] contains either a red
copy of M or a blue copy of the trivially ordered χ-partite graph Kn′,n′,...,n′ with χ = 2j−1. Partition
[Nj ] into n consecutive intervals V1, V2, . . . , Vn, each of length Nj−1. We again try to embed a red
copy of M by placing vertex i in the set Vi for each i = 1, 2, . . . , n. If this procedure does not
produce a red copy of M , then there are two indices i1 and i2 such that every edge between Vi1and Vi2 is blue. Moreover, by the induction hypothesis, either one of Vi1 or Vi2 contains a red copy
of M , in which case we are done, or both Vi1 and Vi2 contain a blue copy of the 2j−1-partite graph
Kn′,n′,...,n′ . Combining the two gives a blue copy of the 2j-partite graph Kn′,n′,...,n′ .
Since any ordered matching M on n vertices is a subgraph of Kn = K1,1,...,1, the n-partite graph
with parts of size 1, we see that r<(M) ≤ r<(M,Kn) ≤ ndlogne, establishing Theorem 1.2. We
will now prove Theorem 1.1 by showing that r<(M) ≥ nc logn/ log logn for almost all orderings. We
need a simple lemma which says that in a randomly ordered matching with n vertices, that is, an
ordered matching chosen uniformly at random from the set of all possible orderings of a matching
with n vertices, any two disjoint intervals of length at least 4√n log n have an edge between them.
Lemma 2.2. Let M be a random matching on vertex set [n]. Then, asymptotically almost surely,
there is an edge between any two disjoint intervals of length at least 4√n log n.
Proof. Given two disjoint sets A and B of order t, where t is even, the probability that a random
matching has no edges between A and B is at most(n− t− 1
n− 1
)(n− t− 3
n− 3
)· · ·(n− 2t+ 1
n− t+ 1
)≤(n− tn
)t/2≤ e−t2/2n,
6
where the first inequality holds since n−t−kn−k < n−t
n for all k > 0 and the second inequality follows
from the fact that 1−x ≤ e−x for all x. Since there are at most n2 pairs of intervals of length t, the
probability that there exist two intervals of length t with no edge between them is at most n2e−t2/2n.
For t ≥ 3√n log n, this tends to zero with n. Therefore, taking t to be an even integer between
3√n log n and 4
√n log n, we see that asymptotically almost surely any two disjoint intervals of
length at least t have an edge between them, completing the proof.
Theorem 2.3. Let M be a random matching on vertex set [n]. Then, asymptotically almost surely,
r<(M) ≥ nlogn/20 log logn.
Proof. By Lemma 2.2, we may assume that in M any two intervals of length at least 4√n log n
have an edge between them. Let s = bn1/4c and suppose that c is a two-coloring of the edges of
the complete graph on vertex set [s] without a monochromatic clique of order log n (such a coloring
exists by Ramsey’s theorem).
Let G0 be the graph with a single vertex. For i ≥ 1, we recursively define edge-colored ordered
complete graphs Gi as follows. Let Gi,1, Gi,2, . . . , Gi,s be vertex disjoint copies of Gi−1. We form
Gi by placing these copies of Gi,j in sequential order, that is, placing the vertices of Gi,j before the
vertices of Gi,j+1, and, for each 1 ≤ j < j′ ≤ s, adding a complete bipartite graph in color c(j, j′)
between the two graphs Gi,j and Gi,j′ .
We prove by induction on i that the graph Gi does not contain a monochromatic copy of a
subgraph of M induced on a subinterval of size at least n3/4 · (8 log n)i (we refer to such subgraphs
as subintervals of M). The claim is trivially true for i = 0 since Gi consists of a single vertex.
Suppose now that the claim has been proved for Gi−1 and, for the sake of contradiction, suppose
that Gi contains a monochromatic subinterval M ′ of M of size at least n3/4 · (8 log n)i. Abusing
notation, we also let M ′ denote this copy in Gi. Without loss of generality, we may assume that
it is a red copy. For j = 1, 2, . . . , s, let Wj = V (Gi,j) ∩ V (M ′) and note that each Wj forms a
subinterval of M . We call a set Wj small if |Wj | < 4√n log n and large otherwise.
There are at most 4√n log n · s ≤ 4n3/4
√log n vertices of M ′ in small sets Wj . Therefore, at
least 12n
3/4 · (8 log n)i vertices lie in large sets Wj . By our assumption, there exists a red edge of
M between every pair of large sets. However, since c does not contain a monochromatic clique of
order log n, there are fewer than log n large sets Wj . Therefore, for some index j, we have
|V (Gi,j) ∩ V (M ′)| ≥ 1
log n· 1
2n3/4(8 log n)i ≥ n3/4 · (8 log n)i−1,
contradicting the induction hypothesis. Since
n3/4 · (8 log n)logn/4 log(8 logn) = n,
we see that for t = b logn4 log(8 logn)c the graph Gt does not contain a monochromatic copy of M . Since
Gt has st vertices and, for n sufficiently large, st ≥ nlogn
20 log logn , the claimed lower bound follows.
The only property of M used in the proof of Theorem 2.3 is that there is at least one edge
between any two disjoint intervals of length 4√n log n. Since it is straightforward to construct
7
such matchings explicitly (see the start of Section 4), we also have explicit examples of ordered
matchings with superlinear ordered Ramsey number.
It follows from Theorem 2.1 that if M has interval chromatic number 2 then r<(M) ≤ n2.
We now show that this is close to best possible. We will construct our matching using the well-
known van der Corput sequence or rather a collection of permutations derived from the van der
Corput sequence. To define these permutations, we take the integers 0, 1, . . . , 2h − 1, write each
as a binary expansion with h digits and reverse its expansion. We call the resulting permutation
π : [2h] → [2h] a van der Corput permutation. For n = 2h, these permutations are known to have
the property that for any pair of intervals I, J ⊆ [n],∣∣∣∣|π(I) ∩ J | − |I||J |n
∣∣∣∣ ≤ C log n, (2)
where C is an absolute constant (see, e.g., [29, Chapter 2]). We note that this is a considerably
smaller discrepancy than one could hope to achieve using a random permutation. We are now
ready to prove Theorem 1.4 in the following form.
Theorem 2.4. There exists a positive constant c such that for all n = 2h there is a matching of
interval chromatic number 2 with 2n vertices satisfying
r<(M) ≥ cn2
log2 n log logn.
Proof. Let π be the van der Corput permutation. We take M to be the ordered perfect matching
with interval chromatic 2 on [2n] which matches i ∈ [n] with n+ π(i).
Let s = n8 logn and t = 8cn
logn log logn , where c is a sufficiently small positive constant. Consider a
random red/blue-coloring χ of the edges of the complete graph with loops on [t]. Let Ii = [(i−1)s+
1, is] for 1 ≤ i ≤ t. Let ψ be the red/blue-coloring of the complete graph on [ts], where every edge
between Ii and Ij is of color χ(i, j). We will show that with positive probability the edge coloring
ψ does not have a monochromatic ordered copy of M and, therefore, r<(M) > ts = cn2
log2 n log logn.
By symmetry between the two colors, it suffices to show that the probability that the red graph
contains an ordered copy of M is less than 1/2. Note that the red graph is a blow-up of the random
graph with loops on [t], where each vertex i is replaced by an interval Ii of length s.
If the coloring ψ contains a red copy of M , then there exists an integer k with 1 ≤ k ≤ t and
partitions [n] = A1∪· · ·∪Ak, [n+ 1, 2n] = Bk ∪Bk+1∪· · ·∪Bt of the vertex set of M into intervals
such that Ai embeds into Ii for all 1 ≤ i ≤ k and Bj embeds into Ij for all k ≤ j ≤ t. Each Ai and
Bj has size at most s and if π(Ai) + n contains an element of Bj , then χ(i, j) must be red. We
note that there are at most(2n+ t
t
)≤(e((4c)−1 log n log logn+ 1)
)t ≤ e9cn/ logn
different partitions of this form. We now fix such a partition.
Let d = 2 log n. Let Ai0 be the (d + 1)th largest interval of the form Ai and Bj0 the (d + 1)th
largest interval of the form Bj . If |Ai0 ||Bj0 | > Cn log n, then every interval Ai with size at least
|Ai0 | and every interval Bj with size at least |Bj0 | has an edge of M between them. Therefore, if
8
the partition gave rise to a monochromatic red copy of M , χ would contain a red complete bipartite
graph with parts of size d. The probability of this event is at most
2−d2
(t
d
)2
< 2−d2 t2d
d!2<
1
4.
Otherwise, we may suppose that |Ai0 ||Bj0 | ≤ Cn log n. As each Ai and each Bj has size at most
s, there are at most 2ds edges of M coming from some Ai of size larger than Ai0 or some Bj of size
larger than |Bj0 |. Therefore, there are at least n − 2ds ≥ n/2 edges of M between those Ai and
Bj with |Ai| ≤ |Ai0 | and |Bj | ≤ |Bj0 | and hence between those Ai and Bj with |Ai||Bj | ≤ Cn log n.
By (2), each such pair Ai, Bj has at most|Ai||Bj |
n +C log n ≤ 2C log n edges between them and thus
there are at least n/22C logn = n
4C logn pairs Ai, Bj for which there is at least one edge of M between
Ai and Bj . But this implies that the coloring χ contains a particular red subgraph with at leastn
4C logn edges, which occurs with probability at most 2−n/4C logn. Since the collection of edges is
determined by the choice of the Ai and Bj and there are at most e9cn/ logn choices for these sets,
we therefore see that the probability the coloring ψ contains a red copy of such a graph is at most
e9cn/ logn2−n/4C logn <1
4,
for c sufficiently small. Hence, we see that the probability ψ contains a red copy of M is at most
1/2, completing the proof.
The ordered Ramsey number of a matching is not always controlled by the interval chromatic
number. For example, despite having interval chromatic number which is linear in n, the ordered
Ramsey number of the matching with edges (1, 2), (3, 4), . . . , (n− 1, n) is clearly linear in n. More
generally, if we know that a matching has bounded bandwidth with respect to the given ordering,
that is, there exists a k such that |i− j| ≤ k for any edge (i, j), then we can show that the ordered
Ramsey number is at most a polynomial whose power is dictated by the bandwidth. The proof of
this fact relies on the next lemma.
Given two ordered graphs G and H, we define their ordered lexicographic product G ·H to be
the graph consisting of t := |H| consecutive ordered copies of G, which we call G1, G2, . . . , Gt, with
all vertices of Gi joined to all vertices of Gj if and only if (i, j) is an edge of H.
Lemma 2.5. For any ordered matching M and any ordered graphs G and H,
r<(M,G ·H) ≤ r<(M,G) · r<(M,H).
Proof. Suppose that we are given a two-coloring of the edges of KN with N = r<(M,G) ·r<(M,H).
Partition [N ] into s := r<(M,H) consecutive intervals V1, V2, . . . , Vs, each of length r<(M,G). We
consider the reduced graph with s vertices v1, v2, . . . , vs, connecting vi and vj in red if there are
any edges between Vi and Vj in red and in blue otherwise, that is, if the bipartite graph between
Vi and Vj is completely blue. If the reduced graph contains a red ordered copy of M then so does
the original graph. We may therefore assume that the reduced graph has a blue ordered copy of H
with vertices vi1 , vi2 , . . . , vit , where t := |H|. This gives vertex sets Vi1 , Vi2 , . . . , Vit such that there
are complete blue bipartite graphs between vertex sets corresponding to edges of H. Since Vij has
9
size r<(M,G), we see that either some Vij contains an ordered red copy of M , in which case we
are done, or every Vij contains an ordered blue copy of G. In the latter case, we may use the blue
edges between pieces to get an ordered blue copy of G ·H, completing the proof.
The result about bandwidth mentioned earlier is now an easy consequence of this lemma. We
say that an ordered graph has bandwidth at most k if |i− j| ≤ k for every edge (i, j). We also write
P kn for the kth power of a path, the ordered graph on [n] where i and j are adjacent if and only
if |i − j| ≤ k. Note that an ordered graph on [n] has bandwidth at most k if and only if it is a
subgraph of P kn .
Theorem 2.6. For any ordered matching M on n vertices and any positive integer k,
r<(M,P kn ) ≤ ndlog ke+2.
In particular, if M is an ordered matching on n vertices with bandwidth at most k, then
r<(M) ≤ ndlog ke+2.
Proof. To begin, note that P kn is a subgraph of Kk · Pn. By Lemma 2.5, it follows that
r<(M,P kn ) ≤ r<(M,Kk) · r<(M,Pn).
Theorem 2.1 implies that r<(M,Kk) ≤ ndlog ke and observation (1) from the introduction implies
that r<(M,Pn) ≤ r<(Kn, Pn) ≤ n2. The result follows.
3 General graphs
In this section, we will prove Theorems 1.3 and 1.5. We begin with Theorem 1.3, which is contained
in the following result.
Theorem 3.1. Let H be an ordered d-degenerate graph on n vertices with maximum degree ∆ and
let K = Kn′,n′,··· ,n′ be a trivially ordered complete χ-partite graph. Let s = dlogχe and D = 8χ2n′.
Then
r<(H,K) ≤ 2s2d+s∆snsDds+1.
In particular, if H is an ordered d-degenerate graph on n vertices with interval chromatic number
χ, then
r<(H) ≤ n32d logχ.
The next lemma is the key technical step in the proof of Theorem 3.1. It says that if a graph
on [N ] does not contain a copy of a particular ordered d-degenerate graph H then it contains an
ordered collection of large subsets with low density between each pair of subsets. In the unordered
case, analogues of this result may be found in [19, 23]. To prove the lemma, we will attempt to
embed H greedily. If this fails, we can show that there must be two large vertex sets, say A and B,
with low density between them. We then repeat this procedure inside the two vertex sets A and
B. If there is no copy of H in A, there will be two large vertex subsets A1 and A2 with low density
10
between them and, similarly, if there is no copy of H in B, there will be two large vertex subsets
B1 and B2 with low density between them. Therefore, we have found four large vertex subsets with
low density between each pair of subsets. Iterating this procedure yields the result below.
Lemma 3.2. Let H be an ordered d-degenerate graph on n vertices with maximum degree ∆.
Suppose that a real number 0 < c < 1 and a positive integer s ≥ 1 are given and that N ≥(2∆n(2sc−1)d
)s. If an ordered graph on vertex set [N ] does not contain an ordered copy of H, then
there exist sets W1,W2, . . . ,W2s ⊂ [N ] such that
(i) for all i, |Wi| ≥ csdN(2sd+1∆n)s
,
(ii) for i < j, all vertices in Wi precede all vertices in Wj,
(iii) for i < j, the density of edges between Wi and Wj is at most c.
Proof. We will prove the statement by induction on s beginning with the base case s = 1. Let
v1, v2, . . . , vn be a d-degenerate ordering of the vertices of H. Partition [N ] into n intervals, each
of length at least N2n . If the required ordering of the vertices of H is vi1 , vi2 , . . . , vin , then we will
label the intervals in order as Vi1 , Vi2 , . . . , Vin .
Consider the following process for embedding an ordered copy of H. We will embed vertices
one at a time following the degenerate ordering, at step t embedding the vertex vt into the set Vt.
To this end, we will try to find, by induction, a sequence of vertices w1, w2, . . . , wt with wi ∈ Vi for
i = 1, 2, . . . , t and sets Vi,t ⊂ Vi for i = t+ 1, . . . , n satisfying the following properties:
1. If i, j ≤ t and vi is adjacent to vj , then wi is adjacent to wj .
2. If j ≤ t < i and vi is adjacent to vj , then wj is adjacent to every vertex in Vi,t.
3. |Vi,t| ≥ cdi,t |Vi| for all i > t, where di,t is the number of neighbors of vi in v1, . . . , vt.
If the process reaches step n, Property 1 implies that mapping vi to wi for each i = 1, 2, . . . , n gives
the required ordered copy of H.
To begin the induction, we let Vi,0 = Vi for all i. The properties stated above are then trivially
satisfied. Suppose now that we have found w1, w2, . . . , wt−1 and Vi,t−1 for all i ≥ t and we wish to
define wt and Vi,t for all i > t. Let It = i > t : vi is adjacent to vt and note that |It| ≤ ∆. We
have
di,t =
di,t−1 + 1 if i ∈ It,
di,t−1 if i /∈ It.
For every i > t with i /∈ It, let Vi,t = Vi,t−1 and note that these sets satisfy Properties 2 and 3
above. For a vertex w ∈ Vt,t−1 and an index i ∈ It, let Vi,t(w) be the set of neighbors of w in Vi,t−1.
If there exists a vertex w ∈ Vt,t−1 such that
|Vi,t(w)| ≥ c|Vi,t−1| ≥ cdi,t−1+1|Vi| = cdi,t |Vi|
for all i ∈ It, then we may set Vi,t = Vi,t(w) for i ∈ It, take wt = w and proceed to the next
step. If this is not the case, then for each vertex w in Vt,t−1 there exists an index i ∈ It for which
11
|Vi,t(w)| < c|Vi,t−1|. By the pigeonhole principle, there exists an index i ∈ It such that there are at
least1
|It|· |Vt,t−1| ≥
1
∆cdt,t−1 |Vt| ≥ cdt,t−1
N
2∆n
vertices in Vt,t−1 which all have at most c|Vi,t−1| neighbors in Vi,t−1. Let W1 be these vertices and
W2 = Vi,t−1. Then
|W1|, |W2| ≥cdN
2∆n.
Relabeling W1 and W2 if necessary, we have found sets W1 and W2 satisfying conditions (i), (ii)
and (iii) of the lemma. That is, if we cannot find an ordered copy of H, we can find sets W1 and W2
satisfying properties (i), (ii), and (iii). In fact, W1 and W2 satisfy the following stronger conditions:
(i’) |W1|, |W2| ≥ cdN2∆n ,
(ii’) all vertices in W1 precede all vertices in W2 or all vertices in W2 precede all vertices in W1,
(iii’) each vertex in W1 has at most c|W2| neighbors in W2.
Suppose now that we are given an integer s ≥ 2 and the claim has been proved for all smaller
values of s. By the stronger form of the case s = 1 with c′ = 2−sc, we can find two sets W1 and W2
such that |W1|, |W2| ≥ cdN2sd+1∆n
and all the vertices in W1 have at most c2s |W2| neighbors in W2. By
applying the inductive hypothesis inside W1, we can find sets W1,1, · · · ,W1,2s−1 such that (i) for all
j, |W1,j | ≥ c(s−1)d|W1|(2(s−1)d+1∆n)s−1 ≥ csdN
(2sd+1∆n)s, (ii) for j < j′, the vertices in W1,j precede the vertices in
W1,j′ , and (iii) for all j and j′, the density of edges between W1,j and W1,j′ is at most c.
For each j = 1, 2, . . . , 2s−1, let X2,j ⊂ W2 be the set of vertices which have at least c|W1,j |neighbors in W1,j . By the degree condition on vertices in W1, we see that
|X2,j | · c|W1,j | ≤ |W1,j | ·c
2s|W2|,
and hence |X2,j | ≤ 12s |W2|. Therefore,
∑2s−1
j=1 |X2,j | ≤ |W2|2 .
Let W ′2 = W2 \⋃2s−1
j=1 X2,j and note that |W ′2| ≥|W2|
2 . All vertices in W ′2 have at most
c|W1,j | neighbors in W1,j for all j = 1, · · · , 2s−1. Now apply the s − 1 case to W ′2 to find sets
W2,1, · · · ,W2,2s−1 . We have, for all j,
|W2,j | ≥c(s−1)d|W ′2|
(2(s−1)d+1∆n)s−1≥ csdN
(2sd+1∆n)s,
and (relabeling W1 and W2, if necessary) one can easily check that the sets W1,1, · · · ,W1,2s−1 ,
W2,1, · · · ,W2,2s−1 satisfy the claimed properties.
Before moving on to the next ingredient, we note a corollary which we will need later on.
Corollary 3.3. Let H be an ordered d-degenerate graph on n vertices. Suppose that a real number
0 < c < 12 is given and that N ≥ (n2c−7d)4 log(1/c). If an ordered graph on [N ] does not contain an
ordered copy of H, then there is a subset of order at least (n2c−7d)−4 log(1/c)N with edge density at
most c.
12
Proof. Let s = dlog(2/c)e ≤ 4 log(1/c). Since
N ≥ (n2c−7d)4 log(1/c) ≥ (2∆n(2s(2c−1))d)s,
we may apply Lemma 3.2 with c replaced by c/2. This gives t ≥ 2/c sets W1, . . . ,Wt satisfying
|W1|, . . . , |Wt| ≥ (n2c−7d)−4 log(1/c)N and such that the density of edges between Wi and Wj for
all i < j is at most c/2. For 1 ≤ i ≤ t, let W ′i be a subset of Wi of cardinality exactly N ′ :=
d(n2c−7d)−4 log(1/c)Ne chosen independently and uniformly at random. A first moment calculation
now shows that there is a collection of subsets W ′1 ⊆ W1, . . . ,W′t ⊆ Wt such that |W ′1| = · · · =
|W ′t | = N ′ and∑
i<j e(W′i ,W
′j) ≤ c
2
(t2
)N ′2. We claim that
⋃ti=1W
′i satisfies the required condition.
To see this, note that the number of edges in this set is at most
c
2
(t
2
)N ′2 + t
(N ′
2
)≤(c
2+
1
t
)(tN ′
2
).
Since t ≥ 2/c, the claim follows.
Lemma 3.2 tells us that if the edges of a graph on vertex set [N ] are two-colored in red and
blue and there is no red copy of a particular d-degenerate ordered graph H then it contains an
ordered collection of large subsets with low red density between each pair of subsets. The next
lemma shows that in this situation the blue graph, which has high density between these subsets,
must contain a large trivially ordered multipartite graph.
Lemma 3.4. Let K = Kn,n,··· ,n be a trivially ordered complete χ-partite graph. If a graph on vertex
set [N ] is such that there exist sets W1,W2, · · · ,Wχ satisfying the following conditions:
(i) for all i, |Wi| ≥ 4χn,
(ii) for i < j, the vertices in Wi precede the vertices in Wj,
(iii) for i < j, the density of non-edges between Wi and Wj is at most 18χ2n
,
then the graph contains a copy of K.
Proof. For each pair of distinct i, j ∈ [χ], define Wi,j ⊂Wi as the set of vertices which have at least1
4χn |Wj | non-neighbors in Wj . By property (iii), we see that
|Wi,j | ·1
4χn|Wj | ≤
1
8χ2n|Wi||Wj |,
and thus |Wi,j | ≤ 12χ |Wi|. For each i, define
W ′i = Wi \(⋃j 6=i
Wi,j
),
and note that |W ′i | ≥ |Wi| −∑
j 6=i |Wi,j | ≥ 12 |Wi| ≥ 2χn. For distinct i, j, each vertex in W ′i has at
most 14χn |Wj | ≤ 1
2χn |W′j | non-neighbors in W ′j .
Let v1, v2, · · · , vχn be the vertices of K ordered as in the trivial ordering and let σ(vi) ∈ [χ]
be the index of the part containing vi. We will embed the vertices of K one at a time. At the
13
t-th step, we map vt to a vertex wt ∈ W ′σ(vt). Note that there are at most χn neighbors of vt
in v1, v2, · · · , vt−1. Each such neighbor can forbid at most 12χn |W
′σ(vt)| vertices from being the
image of wt. Also, there are at most n − 1 vertices in W ′σ(vt)already used for embedded vertices.
Therefore, the number of possible images of vt in W ′σ(vt)is at least
∣∣W ′σ(vt)
∣∣− (n− 1)− χn · 1
2χn|W ′σ(vt)
| ≥ 1
2|W ′σ(vt)
| − (n− 1) ≥ 1.
Thus, we can find a vertex wt ∈W ′σ(vt)for which w1, · · · , wt is a copy of K[v1, · · · , vt]. In the
end, we find a copy of K.
By combining Lemmas 3.2 and 3.4, it is now straightforward to prove Theorem 3.1.
Proof of Theorem 3.1. Suppose that the edges of [N ] have been two-colored in red and blue and
the graph does not contain a red copy of H. Apply Lemma 3.2 with c = 1D , where D = 8χ2n′, and
s = dlogχe to obtain sets W1, · · · ,Wχ such that
(i) for all i, |Wi| ≥ csdN(2sd+1∆n)s
= N
(2sd+1∆nDd)s ≥ D ≥ 4χn′,
(ii) for i < j, the vertices in Wi precede the vertices in Wj , and
(iii) for i < j, the density of red edges between Wi and Wj is at most 1D ≤
18χ2n′ .
We may therefore apply Lemma 3.4 to find a blue copy of K, completing the proof of the first part.
To prove that r<(H) ≤ n32d logχ for any ordered d-degenerate graph on n vertices with interval
chromatic number χ, we may clearly assume that n ≥ 3. SinceH is contained in the trivially ordered
complete χ-partite graph Kn,n,...,n, we may apply the bound from the first part with n′ = n. Noting
that s ≤ 2 logχ and D ≤ 8n3 ≤ n5, we have
r<(H) ≤ 2s2d+s∆snsDds+1 ≤ 22ds2n2sD2ds
≤ χ8d logχn4 logχn20d logχ ≤ n32d logχ,
as required.
We conclude this section by proving Theorem 1.5. We will need the following result of Erdos
and Szemeredi [15], which says that if a graph has low density, then it must contain a larger clique
or independent set than would be guaranteed by Ramsey’s theorem alone.
Lemma 3.5. There exists a positive constant a such that any graph on N vertices with density
c ≤ 1/2 contains a clique or an independent set of order at least a logNc log(1/c) .
Using this lemma and Corollary 3.3, it is now easy to prove the following strengthening of
Theorem 1.5.
Theorem 3.6. There exists a constant C such that if H is an ordered graph on n vertices with
degeneracy d, then
r<(H,Kn) ≤ 2Cd log2(2n/d).
14
Proof. Let c = d/n, a be as in Lemma 3.5 and N = max(n2c−7d)8 log(1/c), 22nc log(1/c)/a. Note
that if c ≥ 1/2, the result follows since r<(H,Kn) ≤ r(Kn,Kn) ≤ 22n ≤ 2Cd log2(2n/d) for a large
enough constant C. We may therefore assume that c < 1/2. Suppose now that the edges of the
complete graph on vertex set [N ] are colored with two colors, say, red and blue. If there is no red
copy of H, Corollary 3.3 tells us that there is a subset of size at least (n2c−7d)−4 log(1/c)N ≥ N1/2
with density at most c in red. But then, by Lemma 3.5, the graph contains either a red or a blue
clique of order at least a logN1/2
c log(1/c) ≥ n. The result follows by noting that N ≤ 2Cd log2(2n/d) for a
constant C depending only on a.
4 Off the diagonal
In this section, we will prove Theorem 1.8, that there exists an ordered matching M such that
r<(M,K3) ≥ n4/3−o(1). We will show that this holds when M is a jumbled matching, which we
define to be an ordered matching on [n] satisfying the following two properties:
• every pair of disjoint intervals, each of order at least 2√n, have at least one edge between
them;
• every pair of disjoint intervals, each of order at most 2√n, have at most 9 edges between
them.
To construct such a matching, let n = t2 be an even positive integer. Partition [n] into t
intervals I1, . . . , It, each of order t. There is an ordered matching M on [n] such that each pair Iiand Ij of distinct intervals has an edge between them. Indeed, this can be easily seen by greedily
picking the edges between the intervals. If A and B are disjoint intervals, each of length at least
2t, then there must be at least one edge between them, since each of A and B must completely
contain one of the intervals I1, . . . , It. Moreover, if A and B each have length at most 2t then
there are at most 9 edges between them, since each of A and B intersect at most 3 of the intervals
I1, . . . , It. Therefore, by considering the largest integer t such that n ≥ t2, Theorem 1.8 follows as
an immediate corollary of Theorem 4.2 below. Before tackling this theorem, we recall the famous
Lovasz local lemma, which is a key tool in the proof (see, e.g., [1]).
Lemma 4.1. Let A1, . . . , An be events in an arbitrary probability space. A directed graph D =
(V,E) on the set of vertices V = [n] is called a dependency digraph for the events A1, . . . , An if,
for each i ∈ [n], the event Ai is mutually independent of all the events Aj : (i, j) /∈ E. Suppose
that D = (V,E) is a dependency digraph for the above events and suppose that there are real
numbers x1, . . . , xn such that 0 ≤ xi < 1 and P(Ai) ≤ xi∏
(i,j)∈E(1 − xj) for all i ∈ [n]. Then
P(⋂ni=1Ai) ≥
∏ni=1(1− xi). In particular, with positive probability no event Ai holds.
With this lemma in hand, we are now ready to prove the main result of this section.
Theorem 4.2. There exists a positive constant c such that if M is a jumbled matching on n
vertices, then
r<(M,K3) ≥ c( n
log n
)4/3.
15
Proof. For notational convenience, we will suppose that M is a jumbled matching with 800n log n
vertices and prove that n4/3 ≤ r<(M,K3) for n sufficiently large.
Let G be a given family of ordered graphs on n2/3 vertices, each with at least 40n log n edges, and
suppose that |G| ≤ en2/3 logn. We claim that there exists an edge coloring c1 of the complete graph
over the vertex set [n2/3] in two colors, red and blue, that satisfies the following list of properties:
• there is no blue triangle;
• there is no graph in G all of whose edges are colored red;
• there is no red clique of order at least 20n1/3 log n.
We will prove this claim later. For now, we will assume that it holds and show how it im-
plies the required result. To this end, let N = n4/3 and partition [N ] into consecutive intervals
V1, V2, · · · , Vn2/3 , each of length n2/3. Let Φ be the set of injective embeddings of M into [N ] that
respect the given order of vertices of M . For φ ∈ Φ, let G(φ) be the graph over the vertex set [n2/3],
where there exists an edge between i and j if and only if there is an edge of M between φ−1(Vi)
and φ−1(Vj). Let
G = G(φ) : φ ∈ Φ, e(G(φ)) ≥ 40n log n.
Since the maps in Φ respect the order of vertices, any G(φ) can be described by the last vertex in
φ−1(Vi) for each i ∈ [n2/3]. Therefore, for n sufficiently large,
|G| ≤(
800n log n+ n2/3
n2/3
)≤(e(800n1/3 log n+ 1)
)n2/3
≤ en2/3 logn.
By the claim above, we can find an edge coloring c1 that satisfies the properties listed above with
respect to the family G. Let c2 be the coloring of the complete graph on [N ] where we color the
edges between Vi and Vj with color c1(i, j) for i, j ∈ [n2/3]. We color all the edges within the sets
Vi red. Note that this coloring contains no blue triangle, since c1 does not contain a blue triangle.
Suppose that for φ ∈ Φ, the graph φ(M) forms a red copy of M . Let Wi = V (φ(M)) ∩ Vi for
each i. Let S ⊂ [n2/3] be the set of indices i for which |Wi| ≤ 2√n, and let L = [n2/3]\S. Note that
for a pair of indices i, j ∈ L, since |Wi|, |Wj | > 2√n and M is a jumbled matching, there exists an
edge of M between Wi and Wj . Thus, c1(i, j) is red and we can conclude that the set L forms a red
clique in c1. Hence, since c1 contains no clique of order 20n1/3 log n, we see that |L| ≤ 20n1/3 log n
and ∣∣∣∣∣⋃i∈L
Wi
∣∣∣∣∣ ≤ |L| · n2/3 ≤ 20n log n.
Deleting all edges of M with an endpoint in Wi for some i ∈ L, we see that there are at least
360n log n red edges within⋃i∈SWi. For two indices i, j ∈ S, there are at most 9 edges of M
between Wi and Wj . Hence, G(φ) has at least 40n log n edges, implying that G(φ) ∈ G. However,
G(φ) must be red in the coloring c1 and this is a contradiction. Therefore, φ(M) cannot form a
red copy of M .
It remains to prove the claim. Recall that G is a given family of graphs on n2/3 vertices with at
least 40n log n edges and |G| ≤ en2/3 logn. Let H be a family of graphs on n2/3 vertices with exactly
16
40n log n edges so that for each G ∈ G, there exists a graph H ∈ H satisfying G ⊇ H. We may
choose H so that |H| ≤ |G|. We seek a coloring satisfying:
• there is no blue triangle;
• there is no graph in H all of whose edges are colored red;
• there is no red clique of order at least 20n1/3 log n.
Consider a random coloring of the complete graph on [n2/3] obtained by coloring each edge red
with probability 1 − 12n1/3 and blue with probability 1
2n1/3 . Let Pi be the events corresponding to
blue triangles, Qi the events corresponding to copies of graphs in H being red, and Ri the events
corresponding to red cliques. Let IP , IQ, IR be the index sets for the events Pi, Qi, Ri, respectively.
By applying the local lemma, we will prove that there exists a coloring where none of the events
Pi, Qi, Ri occur. Note that
|IP | =(n2/3
3
)≤ n2, |IQ| ≤ |H| ≤ |G| ≤ en
2/3 logn, |IR| =(
n2/3
20n1/3 log n
)≤ e20n1/3 log2 n.
Let x = 14n , y = e−2n2/3 logn and z = e−40n1/3 log2 n. For later usage, we note that
y|IQ| = o(1) and z|IR| = o(1).
The parameter used in the local lemma will be x for events Pi, y for events Qi, and z for events
Ri. We now check the conditions of the local lemma.
Event Pi : Since Pi is an event depending on three edges, there are at most 3n2/3 other events
Pj that are adjacent to Pi in the dependency graph. For the events Qj and Rj , we use the trivial
bound |IQ| and |IR| for the number of events that depend on Pi. Hence, for Pi, we have
x∏
j∈IP ,j∼i(1− x)
∏j∈IQ,j∼i
(1− y)∏
j∈IR,j∼i(1− z)
= (1− o(1))xe−x(3n2/3)e−y|IQ|e−z|IR|
= (1− o(1))1
4n≥ 1
8n= P(Pi).
Event Qi : Since Qi is an event depending on 40n log n edges, there are at most 40n5/3 log n events
Pj that are adjacent to Qi in the dependency graph. Hence, for Qi, we have
y∏
j∈IP ,j∼i(1− x)
∏j∈IQ,j∼i
(1− y)∏
j∈IR,j∼i(1− z)
= (1− o(1))ye−x(40n5/3 logn)e−y|IQ|e−z|IR|
= (1− o(1))e−2n2/3 logne−10n2/3 logn
≥ e−20n2/3 logn ≥(
1− 1
2n1/3
)40n logn
= P(Qi).
17
Event Ri : Since Ri is an event depending on(
20n1/3 logn2
)≤ 200n2/3 log2 n edges, there are at
most 200n4/3 log2 n events Pj that are adjacent to Ri in the dependency graph. Hence, for Ri, we
have
z∏
j∈IP ,j∼i(1− x)
∏j∈IQ,j∼i
(1− y)∏
j∈IR,j∼i(1− z)
= (1− o(1))ze−x(200n4/3 log2 n)e−y|IQ|e−z|IR|
= (1− o(1))e−40n1/3 log2 ne−50n1/3 log2 n
≥ e−95n1/3 log2 n ≥(
1− 1
2n1/3
)(20n1/3 logn
2 )= P(Ri).
The result follows.
While we suspect that r<(M,K3) ≤ n2−ε for some ε > 0, we have been unable to improve on
the trivial bound r<(M,K3) ≤ r(Kn,K3) = O(n2/ log n).
5 Odds and ends
5.1 Connection to hypergraphs
Given a 3-uniform hypergraph H, we define the Ramsey number r(H) to be the smallest natural
number N such that every two-coloring of the edges of the complete 3-uniform hypergraph K(3)N
contains a monochromatic copy of H. In this subsection, we will show that for any 3-uniform
hypergraph H, there is a family of ordered graphs SH such that the Ramsey number of H is
bounded in terms of the ordered Ramsey number of the family SH, where the ordered Ramsey
number r<(F) for a family of ordered graphs F is defined to be the smallest N such that every
two-coloring of the edges of 1, 2, . . . , N contains an ordered copy of some F ∈ F .
For any ordered graph H on 1, 2, . . . , n, we define a 3-uniform hypergraph T (H) on vertex set
1, 2, . . . , n+1 by taking all triples whose first pair is an edge of H. Given a 3-uniform hypergraph
H on n+ 1 vertices, we let SH be the collection of ordered graphs H on 1, 2, . . . , n such that H is
a subhypergraph of T (H). For example, if H = K(3)n+1 then SH = Kn and if H = K
(3)4 \e then SH
contains three different graphs on vertex set 1, 2, 3, namely, the complete graph K3 and the two
graphs with two edges containing the edge 1, 2. Note that for any graph H, we have H ∈ ST (H).
Our main theorem relating upper bounds for Ramsey numbers of 3-uniform hypergraphs to ordered
Ramsey numbers is now as follows.
Theorem 5.1. Let H be a 3-uniform hypergraph. Then
r(H) ≤ 2(r<(SH)2 ) + 1.
Proof. The result follows from the method of Erdos and Rado. Suppose that N = 2(r<(SH)2 ) + 1
and the edges of the complete 3-uniform hypergraph on 1, 2, . . . , N have been two-colored, say
by red and blue. We will find an increasing sequence of vertices v1, v2, . . . , vt+1 with t = r<(SH)
18
such that, for any given i and j with i < j, all triples of the form vi, vj , vk with k > j have the
same color χ(i, j). Consider the two-coloring of the edges of the complete graph on v1, v2, . . . , vtwhere the edge vi, vj receives color χ(i, j). Then, since t = r<(SH), this graph must contain a
monochromatic copy of some ordered graph in SH. By construction, the 3-uniform hypergraph on
v1, v2, . . . , vt+1 must then contain a monochromatic copy of H in the same color. It only remains
to find the sequence v1, v2, . . . , vt+1.
We will prove, by induction, that for any 1 ≤ ` ≤ t, there is a sequence of vertices v1, v2, . . . , v`and a set V` with
|V`| ≥ 2(r<(SH)2 )−(`
2)
such that, for any 1 ≤ i < j ≤ `, all triples vi, vj , w with w ∈ vj+1, . . . , v` ∪ V` have the same
color χ(i, j) depending only on i and j.
To begin, we let v1 = 1 and V1 = 2, 3, . . . , N. Suppose now that v1, v2, . . . , v` and V` have
been constructed satisfying the required conditions and we wish to find v`+1 and V`+1. We let v`+1
be the smallest element of V`. Let V`,0 = V` \ v`+1. We will construct a sequence of subsets
V`,0 ⊃ V`,1 ⊃ · · · ⊃ V`,` such that, for each j, all triples vj , v`+1, w with w ∈ V`,j are of the same
color that depends only on the index j. Note that since V`,` ⊂ · · · ⊂ V`,0 ⊂ V` it follows that all
triples vi, vj , w with 1 ≤ i < j ≤ ` + 1 and w ∈ V`,` have the same color depending only on the
values of i and j.
Suppose now that V`,j has been constructed in an appropriate fashion. To construct V`,j+1, we
consider the set of vertices w ∈ V`,j for which vj+1, v`+1, w have color red. If this set has size at
least |V`,j |/2, we let V`,j+1 be this set. Otherwise, we let V`,j+1 be the complement of this set in
V`,j . In either case, |V`,j+1| ≥ |V`,j |/2. To finish the construction of V`+1, we let V`+1 = V`,`. Note
that for each j ≤ ` and every w ∈ V`+1 the triple vj , v`+1, w has the same color. Furthermore,
|V`+1| ≥⌈|V`| − 1
2`
⌉≥ 2(r<(SH)
2 )−(`+12 ).
In particular, |Vt| ≥ 1 and choosing vt+1 ∈ Vt completes the proof.
We expect that the bound coming from this theorem is close to sharp in many cases. For
example, for H = K(3)n+1, the complete 3-uniform hypergraph, we have SH = Kn, so Theorem 5.1
returns the usual double-exponential bound, which is believed to be tight. However, there are cases
where the bound given above is far from the truth. This can be seen by considering the balanced
complete tripartite 3-uniform hypergraph H = K(3)n,n,n. It has Ramsey number 2Θ(n2) but the best
bound the theorem above can give is of the form 22O(n)since all ordered graphs in SH contain the
complete bipartite graph Kn,n as a subgraph, thus implying that r<(SH) ≥ r(Kn,n) = 2Ω(n).
Another interesting example to consider is the hypergraph T (H) defined earlier. Since H ∈ST (H), Theorem 5.1 has the following corollary.
Corollary 5.2. For any ordered graph H,
r(T (H)) ≤ 2(r<(H)2 ) + 1.
19
This implies, for example, that if M is an ordered matching on 1, 2, . . . , n, then the Ramsey
number of the hypergraph T (M) is at most 2n4 logn
. For random matchings, we expect that this
unusual looking bound is not far from the truth. We hope to discuss this direction further in future
work.
5.2 Linear ordered Ramsey numbers
In this subsection, we characterize those graphs for which the ordered Ramsey number is linear in
every ordering. A vertex cover V of a graph G is a set of vertices such that every edge of G has an
endpoint in V . The cover number τ(G) of a graph G is the minimum size of a vertex cover of G.
It is easy to show that the cover number of a graph is within a factor two of the size of the largest
matching in the graph. We will prove the following theorem.
Theorem 5.3. A graph G on n vertices has ordered Ramsey number O(n) in every ordering if and
only if τ(G) = O(1).
We first show that if an ordered graph has cover number O(1), then its ordered Ramsey number
is linear in the number of vertices.
Theorem 5.4. For every positive integer τ , there exists a constant c(τ) such that every ordered
graph G on n vertices with a vertex cover of size τ satisfies
r<(G) ≤ c(τ)n.
Proof. By Ramsey’s theorem and a standard averaging argument, there is an integer A and a > 0
such that, for N ≥ A, every two-coloring of the edges of the complete graph on [N ] contains at least
aN2τ+1 monochromatic copies of K2τ+1. Suppose now that we are given an ordered graph G on n
vertices with a vertex cover of size τ . Let N = maxA, 2an and note that every two-coloring of the
edges of [N ] contains at least aN2τ+1 monochromatic copies of K2τ+1. There exists a collection
of vertices v2, v4, . . . , v2τ such that at least aN τ+1 copies of K2τ+1 have their 2nd, 4th, . . . , (2τ)-th
vertex as v2, v4, . . . , v2τ , in order. Without loss of generality, we may assume that at least half of
these copies are red. We let K be this collection of red copies of K2τ+1, noting that |K| ≥ a2N
τ+1.
We also let v0 = 0 and v2τ+2 = N + 1.
We now show that we can find in this coloring a red copy of every ordered graph G on n vertices
which has a vertex cover of size τ . It follows that r<(G) ≤ N so that we may take c(τ) = max(A, 2a).
For i = 0, . . . , τ , let Xi be the set of vertices v between v2i and v2i+2 such that there exists a K ∈ Kwhich has v as its (2i+ 1)-th vertex. We see that
|K| ≤ |X0| · |X1| · · · · · |Xτ |.
Since |Xi| ≤ N for all i and |K| ≥ a2N
τ+1, we see that |Xi| ≥ a2N ≥ n for all i. Note that the
vertices in Xi are incident to all vertices in v2, v4, . . . , v2τ by a red edge. Since |Xi| ≥ n for all i,
we can find a red copy of G. Indeed, we can use the vertices v2, v4, . . . , v2τ as the τ vertices of the
vertex cover and use common neighbors to embed the remaining vertices in the red copy of G.
20
Ordered Ramsey numbers are monotone. That is, if H is an ordered subgraph of G, then
r<(G) ≥ r<(H). The next theorem improves upon this simple fact if H is a small subgraph of G.
We will abuse notation slightly in the statement and proof by using G and H to refer to both the
unordered graphs and particular ordered versions of these graphs.
Theorem 5.5. Let G be a graph on n vertices and H a subgraph on t vertices. Then, for every
ordering of H, there is an ordering of G such that r<(G) ≥ bnt c(r<(H)− 1).
Proof. Let s = bnt c. Given an ordering of H with vertex set [t], we consider any ordering of G
where vertex i of H is in place 1 + (i− 1)s. This is chosen so that H keeps its ordering within G
and any interval of s vertices has at most one vertex from H.
Let r = r<(H)− 1. By the definition of the ordered Ramsey number, there is an edge coloring
c of the complete ordered graph on vertex set [r] containing no monochromatic copy of H with the
given ordering. We will construct a coloring of the complete ordered graph on N = sr vertices with
no monochromatic copy of G with the ordering above. Partition the complete ordered graph on
vertex set [N ] into r intervals each with s vertices. We color every edge between the ith interval
and the jth interval with color c(i, j) and color the edges inside each interval arbitrarily.
We claim that this edge coloring of [N ] has no monochromatic copy of G with the ordering
defined above. Indeed, if there were such a copy of G, the vertices of the ordered subgraph H
would have to be in distinct intervals. This is because the ordering of G was chosen so that no two
vertices of H are contained in an interval of s vertices. However, if the vertices of the monochromatic
ordered copy of H are in distinct intervals, then we would also get an ordered monochromatic copy
of H in the edge coloring c of [r]. This contradicts the definition of c and completes the proof.
We will now use this result to show that if a graph on n vertices has large cover number, then
there is an ordering for which the ordered Ramsey number is large.
Corollary 5.6. There is a positive constant c such that every graph G with n vertices and cover
number τ has an ordering with ordered Ramsey number at least τ c log τ/ log log τn.
Proof. Since G has cover number τ , it contains a matching with t ≥ τ vertices. Consider the
induced subgraph H of G on these t vertices. By Theorem 1.1, there is an ordering of H with
Ramsey number at least tc log t/ log log t. By Theorem 5.5, there is an ordering of G with Ramsey
number at least bnt c(tc log t/ log log t−1). By making c a little smaller, we obtain the desired result.
If G has n vertices and cover number τ(G) = ω(1), Corollary 5.6 implies that there is an
ordering of G with ordered Ramsey number ω(n). Hence, Theorem 5.3 follows immediately from
Theorem 5.4 and Corollary 5.6.
In fact, as a random ordering of the vertices of a graph will likely space out most of the vertices
of a given matching and a random matching almost surely has superlinear ordered Ramsey number,
a minor modification of the proof above gives the following stronger statement. It shows that if the
cover number of a graph is ω(1), then almost all orderings will have superlinear ordered Ramsey
number. We leave the details to the interested reader.
Proposition 5.7. There is a positive constant c such that if G is a graph with n vertices and cover
number τ , then almost every ordering of G has ordered Ramsey number at least τ c log τ/ log log τn.
21
5.3 Density conditions for ordered matchings
For many of the bounds in this paper, we prove stronger off-diagonal results. For example, when
bounding the ordered Ramsey number of an ordered d-degenerate graphH on n vertices, we actually
estimate r<(H,Kn). When bounding this quantity, we first try to embed H greedily. If this fails,
we show that there must be a large subset where the density of edges is at least 1− 1n , so that we
can easily embed copies of Kn. However, this seems incredibly wasteful for bounding r<(H), given
that unordered copies of H already appear at density roughly 1 − 1d . Somewhat surprisingly, this
intuition is wrong. Indeed, we will now show the existence of a positive constant c for which there
is an ordered matching M on n vertices with interval chromatic number 2 and an ordered graph
G on N = 2nc
vertices with edge density at least 1− n−c which does not contain M as an ordered
subgraph. Note that for any fixed ordered graph H with interval chromatic number 2, the ordered
extremal number ex<(N,H), defined in the introduction, is O(N2−ε). However, even for a typical
ordered matching, this result shows that this subquadratic behavior is only exhibited for rather
large values of N .
Let H be an ordered graph with interval chromatic number 2 and k+ ` vertices with every edge
going between [k] and [k + 1, k + `]. Let B be an ordered graph with interval chromatic number 2
and N +M vertices with every edge going between [N ] and [N + 1, N +M ]. We say that H is an
interval minor of B if there are partitions [N ] = I1 ∪ · · · ∪ Ik and [N + 1, N +M ] = Ik+1 ∪ · · · ∪ Ik+`
into intervals with the vertices in Ii coming before the vertices in Ij for i < j such that if (i, j) is
an edge of H, then there is at least one edge in B between Ii and Ij .
Proposition 5.8. There is an ordered matching M on 2n vertices with interval chromatic number
2 and an ordered graph G on 2Ω(n1/12) vertices with edge density 1−O(n−1/6) which does not contain
M as an ordered subgraph.
Proof. Let M be an ordered matching on 2n vertices with interval chromatic number 2, where each
part is partitioned into n1/2 intervals of size n1/2, with an edge between each interval in the first
part and each interval in the second part. Such a matching can be constructed greedily.
Let k = 12n
1/3− 2. In the paper [17], the second author constructed an ordered bipartite graph
B with parts of size t = 2Ω(k1/4) and density 1 − O(k−1/2) across the two parts which does not
contain Jk as an interval minor, where Jk is the ordered graph of interval chromatic number 2
on k + k vertices whose edge set consists of all k2 pairs (i, j) with 1 ≤ i ≤ k < j ≤ 2k. We
now construct a graph G on N = 2n1/6t = 2Ω(n1/12) vertices. Partition the vertex set into 2n1/6
consecutive intervals I1, . . . , I2n1/6 of t vertices each and place a copy of B between each pair of
intervals, while each of the 2n1/6 intervals forms an independent set. The resulting graph G has
density at least 1− 1/(2n1/6)−O(k−1/2) = 1−O(n−1/6).
If G contains M as an ordered subgraph, then, by the pigeonhole principle, there is an interval
Ii containing at least n/(2n1/6) = n5/6/2 vertices from the first part of M . Recall that the first part
of the matching M was partitioned into n1/2 intervals of size n1/2. Hence, Ii will contain at least12n
5/6/n1/2−2 = k such intervals. Similarly, there is an interval Ij containing at least k intervals of
the second part of M . Therefore, the subgraph induced on Ii ∪ Ij contains Jk as an interval minor.
However, this contradicts our choice of B, completing the proof.
22
5.4 Induced ordered Ramsey numbers
A graph H is said to be an induced subgraph of a graph G if V (H) ⊂ V (G) and two vertices of H
are adjacent if and only if they are adjacent in G. The induced Ramsey number r∗(H) is defined
to be the minimum N for which there is a graph G such that every two-coloring of the edges of G
contains a monochromatic induced copy of H.
That these numbers exist was proved by several groups of authors in the early seventies, though
the original methods did not give good bounds for r∗(H). The first major breakthrough on quan-
titative estimates was due to Kohayakawa, Promel and Rodl [25], who proved that there exists a
constant c such that if H is a graph on n vertices, then r∗(H) ≤ 2cn log2 n (see also [19]). This
result was recently improved in [11] to r∗(H) ≤ 2cn logn, edging closer to resolving the well-known
conjecture of Erdos that r∗(H) ≤ 2cn.
It is also possible to define an ordered version of the induced Ramsey number, which we denote
by r∗<(H). This is the smallest N for which there is an ordered graph G such that every two-
coloring of the edges of G contains a monochromatic ordered induced copy of H. A straightforward
modification of the proof of Theorem 1.2 in [11] allows one to prove a bound for r∗<(H) which brings
it in line with the bounds mentioned above. We omit the details.
Theorem 5.9. There exists a constant c such that for any ordered graph H on n vertices,
r∗<(H) ≤ 2cn logn.
Moreover, by modifying the proof of Theorem 1.4 from [19], we may prove the following induced
variant of Theorem 1.3. We again omit the details.
Theorem 5.10. There exists a constant c such that for any ordered graph H on n vertices with
degeneracy d and interval chromatic number χ,
r∗<(H) ≤ ncd logχ.
6 Open problems
Many difficult and interesting problems arose in our study of ordered Ramsey numbers. In this
section, we would like to draw attention to a few of them.
Our first problem relates to the off-diagonal ordered Ramsey number r<(M,K3). We have
already mentioned the trivial upper bound r<(M,K3) ≤ r(Kn,K3) = O(n2/ log n) for any matching
M on n vertices, while Theorem 1.8 shows that there are matchings M on n vertices such that
r<(M,K3) = Ω((n/ log n)4/3). We are not sure where the truth should lie, though we expect that
the upper bound is far from optimal.
Problem 6.1. Does there exist an ε > 0 such that any ordered matching M on n vertices satisfies
r<(M,K3) = O(n2−ε)?
In Theorem 1.1, we proved that there are matchings M with n vertices for which r<(M) ≥nc logn/ log logn, while Theorem 1.2 showed that this result is not far from the truth, in that r<(M) ≤ndlogne for every ordered matching M on n vertices. It would be very interesting to close the gap
between these two bounds.
23
Problem 6.2. Close the gap between the upper and lower bounds for ordered Ramsey numbers of
matchings.
One of our most elementary observations (see Theorem 2.1) was that if M is a matching on
n vertices with interval chromatic number 2, then r<(M) ≤ n2. On the other hand, Theorem 1.4
shows that there are matchings M with n vertices and interval chromatic number 2 for which
r<(M) = Ω(n2/ log2 n log log n). It would again be interesting to close the gap between these two
bounds.
Problem 6.3. Close the gap between the upper and lower bounds for ordered Ramsey numbers of
matchings with interval chromatic number 2.
The d-dimensional cube is the graph on vertex set 0, 1d where two vertices are connected by
an edge if and only if they differ in exactly one coordinate. This is a d-regular bipartite graph
with 2d vertices. It is a well-known open problem to determine whether the Ramsey number of the
cube is linear in the number of vertices, that is, whether r(Qd) = O(2d). The best known upper
bound [12, 28] is quadratic in the number of vertices.
For any ordering of the cube, Theorem 3.1 easily implies that r<(Qd) ≤ 2cd3, while, for a
random ordering, Theorem 2.3 and the fact that Qd contains a matching of size 2d−1 easily imply
that r<(Qd) ≥ 2c′d2/ log d. We believe that the lower bound is closer to the truth.
Problem 6.4. Improve the upper bound for the ordered Ramsey number of the cube.
For orderings of the cube with interval chromatic number 2, Theorem 3.1 again implies that
r<(Qd) ≤ 2cd2, while Theorem 1.4 implies that there exist orderings such that r<(Qd) ≥ c′ 22d
d2 log d.
Again, it seems more likely that the lower bound is closer to the truth.
Problem 6.5. Improve the upper bound for the ordered Ramsey number of cubes with interval
chromatic number 2.
Using a result of Furedi and Hajnal [22], Balko, Cibulka, Kral and Kyncl [4] noted that there
are orderings of the path on n vertices for which the ordered Ramsey number is linear in n. It would
also be interesting to decide whether a similar phenomenon holds for other graphs, that is, whether
there are orderings where the ordered Ramsey number is close to the usual Ramsey number. In
particular, it would be interesting to decide whether there are orderings of the cube for which the
ordered Ramsey number is as small as the usual Ramsey number.
Problem 6.6. Is there an ordering of the cube such that r<(Qd) ≤ 2cd for some absolute constant
c?
We have characterized those graphs such that for every ordering of a graph H, the ordered
Ramsey number is linear in |H|. These are precisely the graphs H for which the edges of H can be
covered by O(1) vertices. A problem of a similar nature is to determine which graphs have some
ordering for which the ordered Ramsey number is linear. In particular, since the Ramsey number
for graphs of bounded maximum degree is linear in the number of vertices, it is natural to ask
whether there are always orderings of d-regular graphs for which the ordered Ramsey number is
linear. We think this unlikely for random regular graphs.
24
Problem 6.7. Do random 3-regular graphs have superlinear ordered Ramsey numbers for all or-
derings?
We note that if the answer to this question is positive, as expected, it would also show that there
are graphs for which the ordered Ramsey number is always significantly larger than the Ramsey
number, regardless of the ordering.
We only have a rather poor understanding of ordered Ramsey numbers for more than two
colors. For example, we could only prove the bound r<(M ; q) ≤ n(2 logn)q−1for any matching M
on n vertices. We believe that something much stronger should hold.
Problem 6.8. Show that for any natural number q ≥ 3 there exists a constant cq such that
r<(M ; q) ≤ ncq logn for any matching M on n vertices.
Finally, we recall Theorem 2.6, which says that for any natural number k and any ordered
matching M on n vertices with bandwidth at most k, the ordered Ramsey number r<(M) satisfies
r<(M) ≤ ndlog ke+2. It would be very interesting to know whether this theorem can be extended to
graphs other than matchings. That is, we have the following problem.
Problem 6.9. Show that for any natural number k there exists a constant ck such that r<(H) ≤ nckfor any ordered graph H on n vertices with bandwidth at most k.
Note added. Before the final version of this article went to print, a solution to Problem 6.9 was
found by Balko, Cibulka, Kral and Kyncl and added to their paper [4]. Their result follows from an
iterated application of Theorem 5.4. To see this, consider the ordered graph Ht,k on 2t+ k vertices
consisting of three successive intervals L,M and R, with |L| = |R| = t and |M | = k, where M is
a copy of the complete graph Kk and every vertex in L and R is connected to every vertex in M .
Since k is fixed, Theorem 5.4 implies that every two-coloring of the edges of the complete graph on
1, 2, . . . , n contains a monochromatic copy of Ht,k with t ≥ cn, where c > 0 depends only on k.
Letting L′ and R′ be the intervals corresponding to L and R in this copy of Ht,k, we see that we
may again apply Theorem 5.4 to find a monochromatic copy of Ht′,k in each of L′ and R′, where
t′ ≥ ct. Iterating this procedure Ωk(log n) times, it is possible to find a monochromatic subgraph
which contains all graphs with nεk vertices and bandwidth at most k. We omit the details, referring
the interested reader instead to [4].
Acknowledgements. We would like to thank the anonymous referees for a number of helpful
remarks.
References
[1] N. Alon and J. H. Spencer, The probabilistic method, 3rd ed., John Wiley & Sons, Hobo-
ken, NJ, 2008.
[2] N. Alon, M. Krivelevich and B. Sudakov, Turan numbers of bipartite graphs and related
Ramsey-type questions, Combin. Prob. Comput. 12 (2003), 477–494.
25
[3] M. Ajtai, J. Komlos and E. Szemeredi, A note on Ramsey numbers, J. Combin. Theory Ser.
A 29 (1980), 354–360.
[4] M. Balko, J. Cibulka, K. Kral and J. Kyncl, Ramsey numbers of ordered graphs, arXiv preprint,
arXiv:1310.7208 [math.CO].
[5] T. Bohman and P. Keevash, Dynamic concentration of the triangle-free process, arXiv preprint,
arXiv:1302.5963 [math.CO].
[6] S. A. Burr and P. Erdos, On the magnitude of generalized Ramsey numbers for graphs, in
Infinite and Finite Sets, Vol. 1, (Keszthely, 1973), Colloq. Math. Soc. Janos Bolyai, Vol. 10,
214–240, North-Holland, Amsterdam, 1975.
[7] V. Chvatal, V. Rodl, E. Szemeredi and W. T. Trotter, The Ramsey number of a graph with
bounded maximum degree, J. Combin. Theory Ser. B 34 (1983), 239–243.
[8] J. Cibulka, P. Gao, M. Krcal, T. Valla and P. Valtr, On the geometric Ramsey number of
outerplanar graphs, Discrete Comput. Geom. 53 (2015), 64–79.
[9] D. Conlon, A new upper bound for diagonal Ramsey numbers, Ann. of Math. 170 (2009),
941–960.
[10] D. Conlon, The Ramsey number of dense graphs, Bull. London Math. Soc. 45 (2013), 483–496.
[11] D. Conlon, J. Fox and B. Sudakov, On two problems in graph Ramsey theory, Combinatorica
32 (2012), 513–535.
[12] D. Conlon, J. Fox and B. Sudakov, Short proofs of some extremal results II, to appear in J.
Combin. Theory Ser. B.
[13] P. Erdos, Some remarks on the theory of graphs, Bull. Amer. Math. Soc. 53 (1947), 292–294.
[14] P. Erdos and G. Szekeres, A combinatorial problem in geometry, Compos. Math. 2 (1935),
463–470.
[15] P. Erdos and E. Szemeredi, On a Ramsey type theorem, Period. Math. Hungar. 2 (1972),
295–299.
[16] G. Fiz Pontiveros, S. Griffiths and R. Morris, The triangle-free process and R(3, k), arXiv
preprint, arXiv:1302.6279 [math.CO].
[17] J. Fox, Stanley–Wilf limits are typically exponential, to appear in Adv. Math.
[18] J. Fox, J. Pach, B. Sudakov and A. Suk, Erdos–Szekeres-type theorems for monotone paths
and convex bodies, Proc. London Math. Soc. 105 (2012), 953–982.
[19] J. Fox and B. Sudakov, Induced Ramsey-type theorems, Adv. Math. 219 (2008), 1771–1800.
[20] J. Fox and B. Sudakov, Two remarks on the Burr–Erdos conjecture, European J. Combin. 30
(2009), 1630–1645.
26
[21] J. Fox and B. Sudakov, Density theorems for bipartite graphs and related Ramsey-type results,
Combinatorica 29 (2009), 153–196.
[22] Z. Furedi and P. Hajnal, Davenport–Schinzel theory of matrices, Discrete Math. 103 (1992),
233–251.
[23] R. L. Graham, V. Rodl and A. Rucinski, On graphs with linear Ramsey numbers, J. Graph
Theory 35 (2000), 176–192.
[24] J. H. Kim, The Ramsey Number R(3, t) has order of magnitude t2/ log t, Random Structures
Algorithms 7 (1995), 173–207.
[25] Y. Kohayakawa, H. Promel and V. Rodl, Induced Ramsey numbers, Combinatorica 18 (1998),
373–404.
[26] A. Kostochka and V. Rodl, On graphs with small Ramsey numbers II, Combinatorica 24
(2004), 389–401.
[27] A. Kostochka and B. Sudakov, On Ramsey numbers of sparse graphs, Combin. Probab. Com-
put. 12 (2003), 627–641.
[28] C. Lee, Ramsey numbers of degenerate graphs, arXiv:1505.04773 [math.CO].
[29] J. Matousek, Geometric discrepancy: an illustrated guide, Algorithms and Combina-
torics 18, Springer-Verlag, Berlin, 1999.
[30] K. Milans, D. Stolee and D. B. West, Ordered Ramsey theory and track representations of
graphs, J. Comb. 6 (2015), 445–456.
[31] G. Moshkovitz and A. Shapira, Ramsey theory, integer partitions and a new proof of the
Erdos–Szekeres theorem, Adv. Math. 262 (2014), 1107–1129.
[32] J. Pach and G. Tardos, Forbidden paths and cycles in ordered graphs and matrices, Israel J.
Math. 155 (2006), 359–380.
[33] J. Pach and G. Toth, Erdos–Szekeres-type theorems for segments and noncrossing convex sets,
Geom. Dedicata 81 (2000), 1–12.
[34] F. P. Ramsey, On a problem of formal logic, Proc. London Math. Soc. 30 (1930), 264–286.
[35] J. H. Spencer, Ramsey’s theorem – a new lower bound, J. Combin. Theory Ser. A 18 (1975),
108–115.
27
